Prognostic models in acute pulmonary embolism: a systematic review and meta-analysis

Background
Existing prognostic models in acute pulmonary embolism (PE) appear inconsistent, but, if accurate, would help to inform treatment decisions and hospital strategies for patients presenting with a PE. This systematic review assesses the quality of studies done to date, synthesizes their findings, and determines how valid and useful the models are for predicting patient outcomes.
Methods and Findings
We did a comprehensive search using OVID MEDLINE and Embase, and The Cochrane Library, and also searched sources of grey literature. Studies were selected and extracted by independent and blinded assessors. Study quality was assessed using a domain approach. Quantitative analyses were performed at relevant risk cut-off levels for each model for various outcomes and time-points. Model validation and model update studies with and without model construction studies were included. We used random-effects meta-analyses and performed subgroup and sensitivity analyses to generate pooled estimates and 95% confidence interval (95% CI) for the proportion of patients in risk groups and event rates within risk groups. We used the hierarchical summary receiver operating characteristic model to produce pooled estimates of sensitivity and specificity.
44 studies (35,667 patients) were included. Among the 12 identified models, 4 were externally validated or updated in 29 studies and 3 were assessed in impact studies. Pulmonary Embolism Severity Index (PESI) and simplified PESI (sPESI) were the most validated models, and PESI was the only model shown to be useful clinically in a randomised trial. The overall proportion [and 95% CI] of patients in the low-risk group for PESI was 40.9% [40.1 – 41.6]. It was 34.4% [32.6 – 36.1] for sPESI. For PESI, the overall in-hospital, 30-day and 3-month mortality rates were 1.4% [1.2 – 1.8] (9 studies), 2.2% [1.9 – 2.6] (7 studies) and 1% [0.5 – 2] (4 validation studies) in the low-risk group, and 9.2% [8.6 – 9.8], 13.8% [13.1 – 14.5] and 12.3% [7 – 20.6] in the high risk-group. For sPESI, the overall 30-day mortality rate was 1.3% [0.7 – 2.3] (6 studies) in the low-risk group and 10.1% [8.4 – 12] in the high risk-group. The lowest event rate in low-risk group was obtained with PE Prognostic Algorithm (in-hospital mortality: 0.6% [0.4 – 0.9]; 30-day mortality: 1.1% [0.8 – 1.5]) but at the expense of a lower prevalence of patients in this low-risk group (21.6% [21.0 – 22.3]) (4 studies). Results are consistent across derivation and validation samples except in the low-risk group for PESI. The highest overall sensitivity estimates were obtained with the Algorithm for predicting in-hospital mortality (97% [96 – 98]) and 30-day mortality (98% [96 – 99]), followed by PESI for 3-month mortality (95% [82 – 99]) and sPESI for 30-day mortality (92% [85 – 96]). The addition of biomarkers to the model provided inconsistent results across studies. Methodological flaws were more common in construction studies than in impact studies.
Conclusions
Most prognostic models in acute PE provide a high predictive ability to identify patients at low-risk, but seem less helpful for identifying those at high-risk. Future research should seek to identify lower risk groups who would be suitable for outpatient management and higher risk groups appropriate for in-hospital aggressive therapy. In regard to the future implementation of new models or refinements of existing models in practice, these should be validated for performance in the same population and compared with decisions based on clinical judgment, before being assessed for utility and cost-effectiveness.